A Generalized Model for Robust Tensor Factorization With Noise Modeling by Mixture of Gaussians

Area of Science:

• Computer Vision
• Machine Learning
• Data Science

Background:

• Low-rank tensor factorization (LRTF) is crucial for preserving structure in computer vision.
• Traditional methods struggle with complex noise, limiting their effectiveness on real-world data.
• Existing least squares and norm-based methods have limitations with non-Gaussian noise and outliers.

Purpose of the Study:

• To develop an advanced LRTF technique robust to complex noise.
• To integrate noise modeling into a generalized weighted LRTF (GWLRTF) framework.
• To enhance low-dimensional subspace recovery in computer vision applications.

Main Methods:

• Proposed a Mixture of Gaussians (MoG) based GWLRTF (MoG GWLRTF).
• Incorporated CANDECOMP/PARAFAC and Tucker factorization into the MoG GWLRTF.
• Utilized the expectation-maximization framework for parameter updates.

Main Results:

• MoG GWLRTF demonstrated superior performance in handling complex noise.
• Both CANDECOMP/PARAFAC and Tucker versions of MoG GWLRTF showed advantages in various applications.
• The proposed method outperformed competing techniques in extensive experiments.

Conclusions:

• The MoG GWLRTF approach offers a robust solution for noisy data in computer vision.
• Integrating noise modeling significantly improves LRTF performance.
• The developed technique enhances subspace recovery accuracy and applicability.